Huawei says Ascend SuperPoD deployments pass 1,000 sets as the 950 SuperPoD wins a World Internet Conference award
Huawei claims the 1,024-NPU system reaches 45% training compute utilisation, which it calls 1.8 times the industry level, and says more than 40 large models have been natively pre-trained on Ascend; none of the figures are independently benchmarked.
Why it matters
Huawei is selling whole supernodes rather than single chips to close the gap with Nvidia, and the deployment count shows how far that strategy has spread among Chinese internet, carrier, finance, education and energy customers. Earlier this year Huawei said more than 750 sets of the previous Ascend 384 SuperPoD were deployed; the new figure covers the whole SuperPoD line, not only the 950 generation. The first live 950DT cluster went into service in Wuhan after being built in about a month. Watch for independent benchmarks of the utilisation and latency claims.
Key facts
- More than 1,000 Ascend SuperPoD sets deployed (whole line)
- Single node: 1,024 NPUs, 1,300+ interconnect chips, 4,000+ optical modules
- Claimed 45% training compute utilisation
- Claimed inference latency as low as 5 ms
- More than 40 large models natively pre-trained on Ascend
- Award: 2026 World Internet Conference Leading Technology Award